Linear scaling in quadratic term for prox operator, general ADMM

[ryanvolz]

Could put in framework to solve a sort of general prox operator, argmin ( f(x) + lmbda/2*||Ax - v||_2^2 ). This can be solved with special methods for some functions (conjugate gradient for f(x) = L2NormSqHalf?), gradient descent for smooth functions, and proximal gradient descent for proximable functions.

Being able to solve such functions would allow implementation of a general ADMM solver that takes problems of the form:
F(x) + G(z)
s.t. Ax + Bz = c.

Whether this is necessary remains to be seen, but it could be useful if general ADMM is faster or if objectives like F(Ax) are required but there is no way to write the closed form prox operator (see #3).